Home

Awesome

Augmenting genetic algorithms with deep neural networks for exploring the chemical space

This repository contains code for the paper: Augmenting genetic algorithms with deep neural networks for exploring the chemical space.

A video summary of the paper can be found here: https://www.youtube.com/watch?v=9VilhlEXm9w&t=16s

Here is a visualization of molecular progress:

<img align="center" src="./readme_docs/mol_view.gif"/>

Prerequisites

For cloning the repository, please have a look at the Branch Navigator section.

Before running the code, please ensure you have the following:

Please note: that the Synthetic Accesability calculater (i.e. directory SAS_calculator) comes from - https://github.com/EricTing/SAscore.

How to run the code? :

We highly recommend using the following version for running your experiments.

python ./core_GA.py

The following settings can be customized (found at the end of the file 'core_GA.py'):

How are the results saved? :

All the results are savents in the 'results' directory. Our results are saved as (Note: 'i' is the run iteration):

  1. images_generation_0_i:
    Images of the top 100 molecules of each generation. Below each molecule are the Fitness, logP, SA, ring penalty and discriminator scores
  2. results_0_i:
    Each sub-directory is named by the generation. The smile strings (ordered by fitness) and corresponding molecular properties are provided as text files: 'smiles_ordered.txt', 'logP_ordered.txt', 'sas_ordered.txt', 'ringP_ordered.txt', 'discrP_ordered.txt'. Outside the sub-directories is the information about the best molecules of a generation.
  3. saved_models_0_i:
    The trained discriminators after each generation. Please Note: We did not make use of the discriminator predictions in the Fitness for this experiment (beta is set to 0).

Branch Navigator:

The code for this repository is arranged based on the experiments of the paper. Particularly: The code for the paper (arranged by experiment) can be found in the paper_results branch. The experiments are arranged as follows:

Instructions on running the experiments of the paper are provided in the above links. Please note that the code has been parallelized based on the number of CPU cores for quick property evaluations.

To run the code quickly, we recommend the following command:

git clone -b master --single-branch https://github.com/aspuru-guzik-group/GA.git --depth 1

This contains the raw GA code, without any results from the paper. Above is very quick for cloning, and has a small file size.

Due to the large size of the repository, we have created a seperate branch that contains outputs from all the eperiment. For this option, please run (note: this is a 4GB branch, and needs 20mins of cloning time):

git clone --single-branch --branch paper_results https://github.com/akshat998/GA.git

Questions, problems?

Make a github issue 😄. Please be as clear and descriptive as possible. Please feel free to reach out in person: (akshat[DOT]nigam[AT]mail[DOT]utoronto[DOT]ca & pascal[DOT]friederich[AT]kit[DOT]edu)

License

Apache License 2.0